Deep-learning-based Registration of Angiogram and Live Fluoroscopy for Percutaneous Coronary Intervention

Abstract

According to 2019 statistics, cardiac disease is the second leading cause of death in South Korea after cancer. Percutaneous coronary intervention (PCI) is an effective procedure for improving cardiovascular disease, but the contrast agent and X-rays used during the procedure can adversely affect the patient’s health, especially in the case of X-rays, which affects cardiologists who continuously perform the procedure. When inserting a guidewire during PCI, a cardiologist watches two X-ray videos, a cine loop of angiography and a real-time video of live fluoroscopy to navigate the guidewire to the target location while avoiding entering wrong branches of blood vessels. However, watching the two separate X-ray videos increases a cardiologist’s mental efforts and intervention time. Moreover, more contrast agents may need to be injected to validate the correct insertion. To address this problem, dynamic coronary roadmapping (DCR) has been suggested because it provides registered images of two different X-ray images. In this study, a novel patient-specific deep-learning-based approach for DCR that uses only two types of X-ray images without using an electrocardiogram (ECG) was proposed. In the proposed approach, a trained deep convolutional neural network (CNN) generates a registered image from the guidewire images in live fluoroscopy and angiographic sequence from diagnostic coronary angiography. Because the proposed approach can implicitly compensate for both cardiac and respiratory motions, it can be applied to arrhythmia patients with irregular ECG signals, unlike conventional image registration through ECG. To confirm the feasibility of the developed method, a CNN was trained in a patient-specific manner using data obtained from clinical PCI procedures on real patients. Qualitative evaluation of 34 PCI cases, including three irregular heartbeat cases, showed low registration error (26% lower than that of the latest work) and high image quality (more than 30 dB peak signal-to-noise ratio (PSNR)). All registration results were classified as “fit for use” (70.6% “Good” and 29.4% “Acceptable”) by three cardiologists. Furthermore, three irregular beating cases were classified as having “Good” registration quality. Registration error (1.06±0.18 mm) was lower than that of a recent work. The method developed in this study is patient-specific; therefore, a new CNN should be trained for each patient in need of a procedure. At this time, it takes less than 10 min to prepare for the intervention after diagnosis; hence, training of the CNN must be completed within that time. For this, an efficient training method is required, and more efficient training is conducted than the training conducted in the feasibility evaluation through transfer learning and improvement of the CNN. The modified registration model requires only approximately 51.5% of the training time of the baseline. Transfer learning showed acceptable performance of the image quality (> 30 dB of PSNR) and registration error (0.72±0.01 mm) even with 5 min of training. In addition, real-time guidewire segmentation using the existing method was conducted to increase usability. Evaluation results of the real-time guidewire segmentation using the models from transfer learning showed slightly degraded image quality (29.87 dB) and registration error (1.46±1.12 mm). However, the image quality is still close to 30 dB and the registration error is comparable to a recent study. In conclusion, the developed DCR method can provide an acceptable visual feedback for both of regular and irregular heart beating, which was assessed by three cardiologists, and can be prepared in the time of the transition between diagnosis and intervention of the PCI.